TinyML Talks Pakistan - SuperSlash - Unifying Design Space Exploration and Model Compression
Explore a comprehensive methodology for unifying design space exploration and model compression in deep learning accelerators for TinyML applications. Delve into the challenges of deploying Deep Learning models on resource-constrained embedded devices and learn about SuperSlash, an innovative solution that combines Design Space Exploration (DSE) and Model Compression techniques. Discover how SuperSlash estimates off-chip memory access volume overhead, evaluates data reuse strategies, and implements layer fusion to optimize performance. Gain insights into the pruning process guided by a ranking function based on explored off-chip memory access costs. Examine the application of this technique to fit large DNN models on accelerators with limited computational resources, using examples such as MobileNet V1. Engage with a detailed analysis of the extended design space, multilayer fusion, and the impact of these strategies on TinyML implementations.